Traditional observed variable multilevel models for evaluating indirect effects are limited by their inability to quantify measurement and sampling error. They are further restricted by being unable to fully separate within- and between-level effects without bias. Doubly latent models reduce these biases by decomposing the observed within-level indicators into within- and between-level latent components. This decomposition simultaneously controls for both measurement error at each level, and sampling error in the aggregation of individual values that serve as indicators of higher-level constructs. We illustrate the usefulness of this approach in investigating the mediating role of aggressive attitudes in associations between reports of school climate and prevalence of teasing and bullying. Teacher perceived justness and support in schools were both found to be indirectly associated with the prevalence of teasing and bullying through aggressive attitudes, and teacher perceived support was found to be directly associated with teasing and bullying.